DeepSeek Harness = Claude Code for $0

0:00 Deepseek just became the fastest growing repo ever and that's because it's harnessed challenges claude code but the app is free open source and runs any model that you want and in this video

0:13 I'll show you the new capabilities it unlocks exactly how to use it so you can build and design with any model and any tools from $0. So if you haven't already grab that beautiful coffee and let's

0:27 dive straight in. So, let's talk about the giant. Well, on everybody's mind, the deepseek harness. But you might be thinking, what on earth is a harness? Think of it like this. The the model is

0:37 like the brain and the body is the harness. Cuz if you don't understand this concept, you won't understand why this is really important to know right now. All I care about is the output

0:48 getting more, getting better. And I want to show you why I'm talking about this today. So, for example, you have Claude the model and you have Claude code the harness. And just like that, there's

0:56 always the Deep Seek model, the V4 Flash, and then there's the harness, which just dropped, which is completely free to use, and we can use any model inside this harness. And in the short

1:08 time it's been here, it's added 170,000 stars. It's got MIT license, so it's basically just yours to use. And you can do any plugins, chat, tools, UI. It's all completely swappable, including

1:18 changing the harness itself. And by the end of this video, you're not only going to have it running with any model you want to, I'm going to show you how you can connect it to anything. So you can

1:26 actually build things very easily without much work. So you may be thinking, Jack, why should I care about this giant blue well? Well, firstly, you can actually choose any model that you

1:36 want to deep sea claw chat GPT and do that straight away. It is as easy as coming down, selecting something, coming here, and changing the model. Secondly, you can actually change this harness in

1:46 any way that you want. It is just literal code that you can do anything with. The app is completely free. So if you have zero dollars, you can use it whenever, however you want to. It runs

1:57 on your machine, not in the cloud. Um, you can basically wire it into anything you're doing with any connectors and you can even grow your own apps from it if you want to. So the idea is we have this

2:07 one universal harness for every Frontier model. Remember, you got Codeex, you got Claude, and now we have Deep Seek. So, all you're going to do to grab this is literally head over to this website

2:17 right here. And all you're going to do is literally come down here and click on view on GitHub or just literally grab this link from down below. Come down to the code and copy it. In the new chat

2:25 window, just be like, "Hey there, go ahead and download this GitHub repo and open it up for me in a local host. Paste it in." And then Cord will do everything. Now, once you've done that,

2:34 you'll open up a new page that looks a little bit something like this. You can select the models if you want to. Again, just grab the skill I put down below in the second link of the description, and

2:43 you can do that. You can select the effort model. You can do anything essentially you want to. And now it effectively just functions exactly like cloud code, but now it's just a deepse

2:52 harness instead. And remember, we can still use all of the cla code models. And if we want to, we can even use free models from open router letting you build very easily. And the cool thing is

3:02 we can make this harness do anything we want to. So I could say something along the lines of, hey there, I'd like you to build yourself a memory tool. Very straightforward. And send that one off

3:11 like so. And you can see and you get a great breakdown of all of the stuff that it's doing. You get loads of granularity and effectively it's think of it like Claude is like a preset Lego build. This

3:22 is really just giving you all the Legos to build anything that you want to. So if you ever find yourself saying hey I wish my harness or you know Claude always spoke to me in a certain way or I

3:31 wish it always had this kind of sign or you don't like the UI. You can effectively change it to do anything that you want. And as you can see as it's going down here we can allow once.

3:41 Now what's really the importance and significance here? Because technically speaking, we can build anything just in claude, just in codecs. Why bother using this? Well, one of the core reasons for

3:52 this, apart from the fact that it is completely model agnostic, we don't need to fudge things around the background or find capabilities. When you own the harness, effectively you're owning the

4:01 intelligence around the model and you can build it in any direction that you want to. You have full autonomy to do that. And you can for example design really cool stuff. You know we covered

4:12 loads of design stuff on here. This is an example of something that I designed earlier. Again this is plain. This is stretch. All I did is I gave it a example UI design. I said go and build

4:21 this for me. And it went ahead and did that. I thought it was really cool. Again just built during the harness. It just makes building things very simple, very crisp and you can build this in any

4:30 way that you like. Again I will say as well honestly looking at this the the UI that it's got is not that bad actually. Now, one of the core ideas here is that we should be using Deepseek, you can use

4:41 any model that you like with this. So, for example, it is completely model agnostic. So, if there's a new flavor of the month, we can just substitute that model in pretty quickly. Now, here's the

4:51 thing. Deepseek, when you compare its performance to Opus and GT 5.6, it is a step below, but it's almost like, would you like 90 95% of the performance for a fraction of the cost? And by the way, if

5:04 this all sounds like I'm speaking Spanish, I'll put a link down below for the full Claude code masterass that talks about these concepts, but really takes it to a new level on foundations,

5:13 building beautiful websites, power features, memory systems, Hermes agent, building anything, design systems, as well as monetization, and also access to the full beautiful uh Claude code and

5:24 Hermes design operating system, uh the design operating system, and a bazillion other things. If you want to level and get miles ahead of everybody else, I'll put a link down below there for you. so

5:33 you can grab it. Then effectively it goes ahead and it completes my ask and I can see exactly how many tokens that took, how long it was, what that looked like and then I said cool hey open up a

5:42 beautiful HTML and explain it to me and then it came ahead and basically went ahead and built me this here. So AI can now remember a lightweight memory layer of a deep sea harness it can save search

5:52 forget. So you can see these are the hallmarks of GPT 5.6 design. So you get the idea how infinitely configurable this is and kind of just how fun it is. But the problem is we don't just want

6:02 something to throw as a chatbot. We actually want to do this as a tool. And I've seen loads of channels covering this. I haven't yet seen anyone say, "Well, how do I do things with it?"

6:10 Well, the cool thing is we can actually, if we want to, get this to actually integrate with things. So, let's take a use case. Let's say that we want to send emails, right? Well, to do that, you're

6:21 just going to use the same skill that I gave you. Go back over to that chat window and basically say to Claude, "Hey there. Now, I want to go ahead and integrate Zappia into my Deep Sea

6:30 Corners so it can then help me draft emails on my behalf. And this is a good use case. And the reason I use Zappia, by the way, is because it is the authentication layer for many of my

6:39 services. Because think about this, we're connected to DeepSeek, Codeex, Claude, Hermes Agent, your personal apps. If you use a service like Zapia, you can basically connect everything

6:50 once and then it's like the same thing and I can control all the different things. But the one of the big reasons I use it is I can actually connect to Outlook very easily and other platforms

6:58 that I couldn't do in others like school for example where my community is and we have a beautiful time there. By the way, we're doing a meet up in Budapest on Saturday. So if you're seeing this, you

7:07 want to swing by, grab a beautiful coffee, we'll be great to to see you there and have an awesome time. So once you've done that, you're going to shoot then over and I might say something like

7:14 uh let's open up a new session. And again, you can pick the workspaces. I might say, hey there, do me a favor. go over to my Gmail and just draft me uh an email to myself describing how beautiful

7:26 Budapest is as a city like so I want to show you how it go ahead and now look at this you can pick the folder you've got standard mode that's your full coding agent edit files that kind of thing

7:35 you've got PTC mode by the way which is all standard mode capabilities with tools exposed through the code mode SDK so the model can combine multi-step operations in one TypeScript program

7:44 minimal mode and then creator mode this one here is for building basically built for creating custo custom agent presets with all standard mode capabilities plus runtime inspection. Honestly, I just

7:54 find myself using standard mode when I've been using this and it's been absolutely fine. But you've got the configurability if you want to. We pick the model and then we send it off here.

8:02 And as you can see, we now have a skill catalog so go through and check out all of my individual skills, which is fantastic. You can see it's now using a tool call, which is awesome. Drafting

8:10 the email and now completing. So now I have essentially anything you can do in the cloud harness you can bring over to Deep Seek. And just like that, it's now drafted the email. And would you believe

8:19 it? We now have the beauty of Budapest app. But it does lead us on to a really important question about when should we use this versus claude. Well, here's a couple of things just to help you make

8:28 an informed decision on it. So, for example, you want to use C code when or codeex in this example. The output is client facing and the polish actually matters. You're building a website like

8:37 we do on the channel. You want really good design taste baked in cuz remember it's brain plus body. and Claude itself has baked in so many different things into its harness that really makes it

8:49 exceptional beyond just using the model. So you may find with certain design things that you're just not getting as good a result as you do in Claude unfortunately but again you can

8:59 systematize that play with that build that up if you want to but Claude is going to be your reliable option there and if you'd rather pay the maintainer when do you want to use deep sea

9:07 harness? Well the first one is volume work where the price is important. So, say for example, we're doing something high volume. Well, we can just quickly open up this deep sea harness in a

9:16 browser and just get it to work on stuff. And it's just very convenient to have this as the workhorse option. This is really important. The way I want to think about AI is you've got frontier,

9:26 so your uh basically Claude Fable, your GPT 5.6 salt, etc. Then there's a role for something called the workhorse. The workhorse is just something that is like 1% the price, 95% of capability. That's

9:40 your Deep Seek. This is a good use case for this. Just send it all the stuff like that in its own window. If you want any model or a free local one Quan comes out um a brand new model that you want

9:50 to play with, well, let's just rip up this Deep Sea Corners and have a conversation with it. And maybe you want to own an extended tool to yourself. It's all great options. Now, what's also

9:59 really good to know is that it is still early, so don't expect it to be perfect just yet, so it may break underneath you. The harness is free, but obviously you have to pay for tokens. Open Router

10:09 itself has a kind of quoteunquote free version. Basically, it just flips between various different free models, but I would recommend honestly just sticking with Deepc FI4 Flash because it

10:18 is just a better model. And as you can see by volume, actually the largest on open router with over 11 trillion tokens. Very, very impressive. Almost as many tokens, guys, is I spend figuring

10:28 out how to get the beautiful design setup you see right behind me there. Not quite that many tokens, maybe a little bit more. Now, a couple of questions you may be thinking about. I wanted to pull

10:35 up here. Number one, is it actually free? Yes. you an MIT license, which is awesome. So, effectively, you are actually just paying $0. You're only paying for tokens. Secondly, does my

10:44 code go to China? Well, no. The tele on your machine, your telemetry is off by default, but you can use a local uh Western model if you really want to make sure. But basically, the telemetry is

10:54 off, so it shouldn't be going to China at all. Can I use my cloud codec subscription with it? No, it's just going to be your API keys. Um, does it work with local models? Yes. Anything

11:03 you use with O Lama, any models like that, you can do it. completely configurable if you want to. And the other one is around plugins. So you can create plugins for this and there is a

11:13 plug-in deepseek store, but honestly I'm not really playing around with the plugins so much. I'm kind of treating is my own thing that I just design and build and have a great time with. Now

11:22 choosing the right model and harness is one thing, but if you don't understand the systems and how to build aic systems that really help you unlock all capabilities, you're leaving too much

11:33 value on the table. So, the next thing we're going to do is learn exactly how to do that in this video right